EG-SPXNet: Edge-Gated Superpixel Graph Neural Networks for Interpretable Retinal Disease Grading

Author First name, Last name, Institution

• Mohamed Elsharkawy, J.B. Speed School of Engineering
• Sadman Sakib, J.B. Speed School of Engineering
• Moumen El-Melegy, J.B. Speed School of Engineering
• Asem Ali, J.B. Speed School of Engineering
• Ali Mahmoud, J.B. Speed School of Engineering
• Mohammed Ghazal, Abu Dhabi University
• Ashraf Khalil, Zayed University, Abu Dhabi CampusFollow
• Wei Wang, University of Louisville School of Medicine
• Ayman El-Baz, J.B. Speed School of Engineering

Document Type

Conference Proceeding

Source of Publication

Lecture Notes in Computer Science

Publication Date

8-5-2026

Abstract

Accurate staging of Age-related Macular Degeneration (AMD) requires detecting and differentiating characteristic lesions: drusen deposits indicating Intermediate AMD, regions of retinal pigment epithelium loss defining Geographic Atrophy, and neovascular membranes with associated exudation characterizing Wet AMD. Conventional deep networks process fundus images as pixel grids without explicitly modeling the spatial relationships between these pathological structures and surrounding healthy tissue. We propose EG-SPXNet, a graph neural network that represents fundus images as superpixel graphs where nodes encode regional features aligned with tissue boundaries and edges capture spatial adjacency between anatomical structures. The central contribution is a learnable edge-gating mechanism that dynamically modulates message passing based on pairwise node relationships, enabling selective amplification of connections between lesion regions and their surroundings while suppressing uninformative pathways. This design provides inherent interpretability through attention maps that highlight drusen clusters, atrophic boundaries, and neovascular complexes in anatomically grounded terms. We evaluate EG-SPXNet on four-class AMD staging, achieving 98.08% accuracy [95% CI: 96.17–99.62], 0.9807 macro F1-score [95% CI: 96.18–99.61], and Cohen’s kappa of 0.9743 [95% CI: 94.88–99.49]. On the ODIR benchmark for binary AMD detection with patient-level splitting, we obtain 91.08% accuracy [95% CI: 88.19–93.70] and 95.66% sensitivity [95% CI: 93.20–97.76]. Our 12.3M parameter model outperforms architectures exceeding 86M parameters under identical evaluation protocols.

DOI Link

10.1007/978-3-032-31438-3_40

ISBN

[9783032314376]

ISSN

0302-9743

Publisher

Springer Nature Switzerland

Volume

16821 LNCS

First Page

605

Last Page

620

Disciplines

Computer Sciences

Keywords

Age-related macular degeneration (AMD), Edge-gated message passing, Fundus images

Scopus ID

105047682557

Recommended Citation

Elsharkawy, Mohamed; Sakib, Sadman; El-Melegy, Moumen; Ali, Asem; Mahmoud, Ali; Ghazal, Mohammed; Khalil, Ashraf; Wang, Wei; and El-Baz, Ayman, "EG-SPXNet: Edge-Gated Superpixel Graph Neural Networks for Interpretable Retinal Disease Grading" (2026). All Works. 8037.
https://zuscholars.zu.ac.ae/works/8037

Indexed in Scopus

yes

Open Access

no